On the Prediction of EPON Traffic Using Polynomial Fitting in
Optical Network Units
I. Mamounakis
1,3
, K. Yiannopoulos
2
, G. Papadimitriou
4
and E. Varvarigos
1,3
1
Computer Technology Institute and Press "Diophantus", Patras, Greece
2
Department of Telecommunications Science and Technology, University of Peloponnese, Peloponnese, Greece
3
Computer Engineering and Informatics Department, University of Patras, Patras, Greece
4
Department of Informatics, Aristotle University of Thessaloniki, Thessaloniki, Greece
Keywords: Prediction, Ethernet Passive Optical Network, Polynomial Prediction, Dynamic Bandwidth Allocation,
Delay.
Abstract: We propose a traffic prediction algorithm that reduces the packets delay in Ethernet Passive Optical
Networks (EPONs). The algorithm relies on Multi-Point Control Protocol (MPCP) message and traffic
monitoring at the Optical Network Units (ONUs) and utilizes the monitoring information to predict the
accumulated burst size using higher order least-mean-square polynomial approximations. The simulation of
the algorithm shows that it achieves a delay improvement of over 30% without any further modification in
the communication and bandwidth assignment procedure of the EPON.
1 INTRODUCTION
Passive optical networks (PONs) (Mukherjee, 2006;
Jason and Vinod, 2005) are an attractive solution for
the deployment of next-generation access networks,
due to their low implementation cost, simple
operation and high-line rates made possible by the
capacity of optical fibers. Ethernet Passive Optical
Networks (EPONs) in particular, which represent the
convergence of low-cost Ethernet equipment and
low-cost fiber infrastructure, find widespread
application in local and metro area networks,
supporting the fiber infrastructure that is being
installed within the scope of fiber-to-the-home,
building and curb (FTTC) end-user access. In an
EPON network, multiple optical network units
(ONUs) access the shared channel to reach the
optical line terminal (OLT) through a passive optical
splitter. To arbitrate the multiple ONU accesses, an
effective bandwidth allocation scheme is required.
The interleaved polling scheme with adaptive cycle
time (IPACT) (Kramer et al., 2002; Luo and Ansari,
2005) is implemented at the OLT. IPACT
periodically receives bandwidth requests from all
connected ONUs and allocate transmission slots
accordingly. The average cycle time in IPACT
contributes to the PON system latency, since ONUs
are served in a round-robin fashion and each ONU
must wait for the full cycle duration before being
served again. Thus, the average cycle time and
consequently the delay, depends on the bandwidth
allocation scheme implemented by IPACT. In
general, bandwidth allocation schemes can be
categorized as fixed or dynamic. Fixed bandwidth
allocation (FBA) schemes (Kramer et al, 2001)
utilize equal size time-slots and offer a fixed time-
slot to each ONU irrespective of its traffic load. The
ONU-to-OLT (upstream) communication channel is
therefore reserved even when the actual ONU traffic
is not sufficient to fully utilize the slot and this
bandwidth underutilization leads to transmission
gaps and increased frame service times. On the other
hand, dynamic bandwidth allocation (DBA) (Kramer
et al., 2003; Mcgarry et al.,2008) assigns the
bandwidth in an adaptive fashion based on the
current traffic load of each ONU. The idea in DBA
schemes, such as the one implemented in IPACT, is
to re-distribute bandwidth from light-load to heavy-
load ONUs within a single cycle duration and
consequently fully utilize the available capacity, thus
reducing the overall PON latency.
An improvement on the PON latency can be
obtained by means of traffic prediction, a technique
that has been widely studied in both wireline and
wireless networks (Zhu, 2012). During the time of
15
Mamounakis I., Yiannopoulos K., Papadimitriou G. and Varvarigos E..
On the Prediction of EPON Traffic Using Polynomial Fitting in Optical Network Units.
DOI: 10.5220/0005056200150021
In Proceedings of the 5th International Conference on Optical Communication Systems (OPTICS-2014), pages 15-21
ISBN: 978-989-758-044-4
Copyright
c
2014 SCITEPRESS (Science and Technology Publications, Lda.)
bandwidth negotiation, and in particular during the
interval that lasts from the moment the ONU sends
the bandwidth requirement until the moment it can
start sending the buffered data. More data will arrive
at the ONU buffer that time and will remain in the
buffer until the next cycle. As a result these will not
be taken into account when the bandwidth request
was sent. OLT-based traffic prediction relies on
estimating future "on-average" bandwidth
requirements for all ONUs in the network based on
their previous bandwidth requests. A key drawback
of OLT-based prediction, however, is that it may not
accurately identify, and therefore respond, to rapid
changes in the ONU traffic. On the other hand
ONU-based prediction, can be performed within a
single cycle, since ONUs are able to constantly
monitor incoming traffic, and therefore can adapt to
traffic changes significantly faster.
Predictive techniques establish a mathematical
model that processes the series of data packets in
order to estimate the future traffic flow. A large
variety of traffic prediction algorithms for EPONs
have been proposed in the last years, in order to
improve the bandwidth allocation strategy and the
total system performance (Mcgarry et al., 2008;
Sadek and Khotanzad, 2004). These prediction
techniques can be executed at the side of OLT
(Hwand et al., 2008; Hwang et al., 2012)
or the
ONUs (Luo and Ansari, 2005; Swades et al., 2010;
Morato et al., 2001; Chan et al., 2009)
with the pros
and cons of each approach that have been described
earlier. The technique proposed in (Hwang et al.,
2008) consists of a two-stage bandwidth request
scheme. In the first stage, DBA is performed for the
next cycle at the ONU level assigning bandwidth to
the ONUs that have more unstable (difficult to
predict) traffic. In this way it becomes easier to
reduce the prediction error by shortening their
waiting times. In the next stage, a linear prediction-
based excess bandwidth request is done for the more
stable ONUs. At the OLT, the proportionally
available bandwidth for an ONU is allocated to
related traffic classes, strictly based on their
respective requests ordered by their priority. In
(Hwang et al., 2008), the authors propose a
prediction process that is based on genetic
expression programming to reduce the queue size
variation and the packet delay. Taking a different
approach (Luo and Ansari, 2005; Swades et al.,
2010; Morato et al., 2001; Chan et al., 2009) propose
prediction techniques that are applied at the ONUs.
In
(Luo and Ansari, 2005), a limited sharing with
traffic prediction scheme was proposed and shown
to enhance DBA process. For ONU-based traffic
prediction another approach was presented in
(Swades et al., 2010) where authors propose a linear
class-based prediction model that tries to estimate
the incoming traffic until the next polling cycle. This
model uses information from previous bandwidth
requests in order to predict bandwidth request at
each ONU in the network, according to the OLT
priority classes. The effect of long-range dependence
of internet traffic in the prediction was studied in
(Morato et al., 2001). In (Chan et al., 2009)
authors
propose a different approach in the EPON by
applying a remote repeater node (RN). The RN
provides electrical regeneration in order to increase
the total number of supported ONU’s and the reach
of the EPON. The proposed scheme processes the
incoming frames in order to improve performance of
downstream and upstream transmissions. While the
prior works have used complicated prediction
techniques at the ONUs, the estimates they produce
refer to a single parameter, that is the bandwidth to
be allocated, which is however a complex metric
(ratio of data size over time duration).
Within the context of ONU-based traffic
prediction, we propose a novel algorithm for
decreasing latency in EPONs. Our algorithm (a)
approximates the frame arrivals within the duration
of a single EPON cycle using least-mean-square
polynomials and (b) estimates the duration of the
upcoming cycle via a least-means-squares adaptive
filter. Subsequently, the two quantities are combined
to produce the amount of data that the ONU will
have accumulated by the time the next bandwidth
assignment from the OLT (GATE message) arrives.
The ONU then communicates the predicted rather
than the actual data to the OLT in the REPORT
message), thus providing the DBA mechanism with
a more informed guess of its traffic requirements.
We show via simulation that the incorporation of the
proposed prediction methods in the EPON operation
can reduce the frame delay from 25% up to 30%
when compared to the standard operation of the
limited and gated versions of Interleaved Polling
with Adaptive Cycle Time (IPACT), depending on
the traffic load and the burstiness of the incoming
traffic. Moreover, this significant performance
benefit is obtained by applying the prediction
algorithms locally at the ONUs and without any
further modification on the Multi-Point Control
Protocol
(MPCP) procedures or the operation of
IPACT. At the same time the proposed solution
exhibits a low computational complexity, which is a
particularly appealing feature when considering the
ONU processing capabilities and associated cost.
The rest of the paper is structured as follows:
OPTICS2014-InternationalConferenceonOpticalCommunicationSystems
16
Section 2 presents our proposed traffic prediction
technique and its scope of application in EPONs.
Section 3 details the simulation setup that was
utilized to evaluate the performance of the prediction
method. Section 4 discusses the results that have
been obtained in terms of latency. Finally, Section 5
concludes the main contributions of this paper.
2 PREDICTION ALGORITHM
BASED ON ONU
In the standard EPON operation, the commu-
nication between the OLT and the ONUs takes place
by means of an interleaved polling scheme with
variable cycle time (IPACT). IPACT operates in
successive cycles, and during each respective cycle
the OLT sends GATE messages that carry
bandwidth grants to all ONUs in the EPON. The
ONUs respond to the GATE messages and send
their data in a co-ordinated fashion, as specified in
the GATE messages, so as to achieve collision free
transmissions in the upstream direction. In addition
to their data, the ONUs also inform the OLT about
their bandwidth requirements (buffer sizes) via
REPORT messages and the IPACT cycle ends upon
the reception of the REPORT messages from all
ONUs in the EPON. At that time, the OLT executes
a dynamic bandwidth allocation (DBA) algorithm to
calculate the grants of the next cycle, and a new
exchange of GATE and REPORT messages ensues.
As a result, the DBA does not take into account (a)
data that have been accumulated at ONUs that are
served near the beginning of the cycle and are forced
to report early, or (b) data that will be accumulated
at ONUs that are served towards the end of the
upcoming cycle and will receive a late grant. This
leads to an additional delay of a cycle time, which
can be particularly significant especially in IPACT
variations with increased or infinite maximum cycle
durations.
The additional delay can be reduced in a
straightforward manner by having each ONU
perform a prediction exactly before the generation of
the current REPORT message by estimating its
buffer occupancy for the instant it will receive the
next GATE message. The ONU can then use the
REPORT message to communicate the prediction to
the OLT rather than the actual (current) buffer size.
Our proposed prediction algorithm of the ONU
buffer size can be summarized as follows:
• Step 1: Constantly monitor the incoming traffic
from hosts in a log file until a GATE message
has been received from the OLT.
• Step 2: Upon the reception of the GATE
message keep a record of its arrival time T(n-1).
• Step 3: Utilize the traffic log to estimate the
instantaneous buffer size B(t).
• Step 4: Utilize the arrival times of previous
GATE messages to predict the arrival time of the
next GATE message T(n).
• Step 4: Combine B(t) and T(n) to calculate the
expected buffer size B(n) at the reception of the
next GATE message.
• Step 4: Transmit the allocated number of frames
in the received GATE and then issue a REPORT
message that carries the bandwidth request B(n).
• Step 5: Reset the traffic log to the remaining
buffer size and re-start from Step 1.
The presented algorithm requires the estimation
of two key parameters: (a) the instantaneous ONU
buffer size B(t), and (b) the arrival time of the next
GATE message T(n+1). The estimation of the
instantaneous buffer size is performed by monitoring
the incoming frames that arrive between REPORT
messages. To this end, the ONU creates a log of the
frame size S
i
and the arrival time t
i
for each frame
that is received. Each frame arrival corresponds to
an increase of the number of bytes B
i
that are stored
at the ONU buffer, following:
,
1 iii
SBB
(1)
while the remaining queue size B
0
after the ONU
transmission at t
0
is used to initialize Eq. (1). Given
(1), a k
th
degree polynomial equation that correlates
the buffer size B(t) and the elapsed time t is can be
calculated by the (t
i
, B
i
) pairs, according to:
02
012
...
k
k
Bt at at at at
(2)
where the coefficients a
0
, a
1,…,
a
k
in the above
polynomial are calculated in a least-mean-squares
fashion by:
1
00
11
1
1....
1...
.... ....
......
1...
k
k
k
kn
n
t
aB
aB
t
aB
t
or
*A TB
(3)
The ONU is able to predict its queue status at
any given future time t and up to the next GATE
message. The exact arrival time of the next GATE
message, however, is not known when the ONU
creates the REPORT message and as a result the
ONU has to estimate it, as well.
OnthePredictionofEPONTrafficUsingPolynomialFittinginOpticalNetworkUnits
17
Table 1: Simulation Parameters.
Symbol Description
Value
(Limited - IPACT)
N
ON
U
Number of ONU’s 8
Physical Layer
Parameters
N
host
Number of ONU
Hosts
15
d ONU distance 10 km
R
d
Downstream Line
Rate
10 Gb/s
R
u
Upstream Line Rate 1 Gb/s
R
n
Host Line Rate 100 Mb/s
IPACT
Parameters
T
max
Max Cycle Time
2 ms Unlimited
W
max
Maximum Grant
Size
82.500 bytes Unlimited
Traffic
Parameters
a (a
O
N
, a
OFF
) Pareto parameter 1.2, 1.5, 1.8
b (b
ON
, b
OFF
) Pareto parameter
a b
O
N
b
OFF
1.2 0,00000375 0,00014 -0,000495
1.5 0,00000375 0,001493-0,000493
1.8 0,00000375 0,0016-0,00048
Prediction
Parameters
p NLMS order 25
M
NLMS step size
constant
0,0001
To this end, the ONU monitors the arrival times
of GATE messages and predicts the arrival time of
the next GATE T(n) by means of a normalized least-
mean-square (NLMS) prediction filter which is
given by:
p
n
i1
ˆ
Tn w i Tn i
(4)
where p is the filter order and w
n
(i) are the filter co-
efficients that are updated at every cycle.
1
2
1
1, 1,,,
ˆ
111,
nn
p
k
Tn i
wi w i Men i p
Tn k
en Tn Tn
(5)
The NLMS step size M has a constant numeric
value (Table 1).
3 SIMULATION SETUP
The performance of our proposed algorithm was
verified via simulation experiments using the
OMNET++ open source simulator (omnetpp.org). In
our setup, a standard EPON architecture
interconnected an OLT with eight ONUs at distances
of 10 km, while the EPON rates were considered
asymmetric (10 Gb/s downstream - 1 Gb/s
upstream). MPCP protocol forms a type of master-
slave REPORT/GATE mechanism, which means
that requirements are put forward by each ONU and
are arbitrated by the optical line terminal (OLT). The
communication model was based on existing
OMNET++ models that provide the basic MPCP
functionalities at the OLT and ONUs. Two IPACT
allocation schemes were implemented at the OLT,
namely, the limited and the gated version.
For the limited-IPACT implementation, OLT
grants an upper bounded transmission window size
per ONU. On the other hand, in case of gated-
IPACT, OLT allocates the estimated requested
bandwidth for each ONU in our network. The
incoming traffic for the purpose of the simulation
was fed to each ONU from an optical switch that
aggregated frames from fifteen independent hosts
(sources) Figure 1. The hosts transmitted data in the
form of fixed size 1000 byte Ethernet frames at a
line rate R
n
of 100 Mb/s. Each of the hosts generated
data frames independently of each other, according
OPTICS2014-InternationalConferenceonOpticalCommunicationSystems
18
Figure 1: Simulation model.
to an ON/OFF traffic model. This traffic model also
known as Pareto distribution consist of two different
periods for each host, an ON (busy) and an OFF
(idle) period. Due to the form of the Pareto
distribution, ON (busy) periods were always
followed by OFF (idle) periods.
The mathematical formula of the distribution is
described in Eq. 6:
1
a
a
x
ba
xf
(6)
Parameters a and b relate to the average busy
time durations and idle time durations, T
ON
and T
OFF
respectively (eq. 7).
.
1
,
1
OFF
OFFOFF
OFF
ON
ONON
ON
a
ba
T
a
ba
T
(7)
The values that were used in our simulations for
the parameters a
ON
, b
ON
and a
OFF
, b
OFF
of the ON
and OFF periods, respectively, are presented in
Table 1. These values resulted in ON-OFF periods
with durations at the msec time scale, which
corresponds to a single IPACT cycle, since an
access-oriented PON is not expected to remain idle
for several successive IPACT cycles. Given the
above average busy and idle periods of each host, it
was possible to calculate the offered loads ρ in the
PON from the number of ONUs (N
ONU
), the number
of hosts per ONU (N
host
) and the individual host load
(ρ
host
) as :
ON
ONU host host ONU host
ON OFF
T
NN NN
TT
(8)
4 RESULTS
We have conducted two sets of simulations
experiments. The first set compares the
limited
IPACT algorithm without prediction to our
corresponding scheme that uses prediction
algorithms. The second set of simulations evaluates
the performance of gated-IPACT algorithm without
prediction against our prediction algorithms. For the
purposes of the simulation, three different traffic
burstiness scenarios that correspond to low burst,
medium burst and high burst (a = 1.8, 1.5, 1.2,
respectively) were used. Moreover, we evaluated the
prediction algorithm for polynomials of degree equal
to one (i.e., linear prediction) and two, since higher
degree for the polynomials lead to severe prediction
inaccuracies that negatively affected the EPON
performance. The respective results are shown in
Figure 2-4 for the limited-IPACT and Figure 5-7 for
the gated-IPACT.
The results clearly demonstrate that the Limited-
IPACT performs in a superior fashion when
prediction based reports are sent by the ONUs. A
percentile delay reduction of over 25% is observed
for medium offered loads around 0.6, while a
smaller benefit is observed as the load becomes
lighter.
For higher loads, prediction only has a minor
beneficial impact when the traffic is relatively
smooth (a=1.8 and 1.5). As the traffic becomes
significantly bursty (a=1.2), the proposed linear
prediction algorithm can be detrimental in terms of
latency, mainly because the cycle durations become
irregular and the GRANT arrival times are not
correctly calculated by the NLMS. As a result,
ONUs request the largest possible grant and IPACT
performs in a TDMA manner with maximum
duration bandwidth grants. For quadratic prediction
the delay results improve in all cases, even for a
highly bursty traffic profile (Figure 4) and the delay
reduction is improved by up to 30% for medium
offered loads around 0.6. As the load increases, the
prediction benefit reduces to under 10%; still, it is
important to notice that quadratic prediction tends to
correct the detrimental effect of liner prediction with
increasing burstiness. A similar behavior is observed
for gated-IPACT in Figure 5-7; the proposed linear
prediction mechanism improves the average delay in
this IPACT variation by 25% for medium loads as
shown in the simulation results. An important
difference with limited-IPACT, however, is
becoming evident for bursty traffic (a=1.2) and at
heavy loads; in this regime even more extended
bandwidth grants are requested by the ONUs and are
allowed by the OLT, due to the fact that gated-
OnthePredictionofEPONTrafficUsingPolynomialFittinginOpticalNetworkUnits
19
Figure 2: Mean Delay for low burstiness traffic.
Figure 3: Mean Delay for medium burstiness traffic.
Figure 4: Mean Delay for high burstiness traffic.
Figure 5: Mean Delay for low burstiness traffic.
Figure 6: Mean Delay for medium burstiness traffic.
Figure 7: Mean Delay for high burstiness traffic.
IPACT does not pose an upper limit on the size of
the grants. As a result, the average delay is also
increased by a significant factor. An even better
performance for the case of gated-IPACT is
observed for quadratic prediction. As it can be seen
from the delay results, an improvement of 26% can
be achieved in medium offered loads from 0.5 to 0.7
for all degrees of traffic burstiness. Moreover, when
the offered loads increase, the utilization of second
order polynomials provides a better delay
performance from its linear counterpart. Especially
for medium burst traffic (a=1.5) the delay is able to
achieve profits up to 27%, around 0.8. Finally, in
accordance with the limited-IPACT results quadratic
prediction algorithm exhibits better stability at high
loads.
5 CONCLUSION
We presented an ONU based prediction method that
is applicable in EPONs. The method relies on the
application of polynomial fitting and the Normalized
OPTICS2014-InternationalConferenceonOpticalCommunicationSystems
20
Least Mean Square (NLMS) algorithms for the
estimation of the instantaneous ONU load and
IPACT cycle duration, respectively, to predict the
ONU buffer size at the time of its next transmission.
We showed via simulations that if the predicted
(estimated) buffer size, rather than the actual size, is
reported to the OLT then a significant (over 25%)
average delay reduction can be realized over
standard EPON when a linear based prediction
algorithm is used. Also when the prediction method
uses a second order polynomial (nonlinear
prediction algorithm) the average delay
improvement is over 30% for all degrees of traffic
burstiness. Moreover,
the proposed techniques are
totally compatible with the bandwidth reporting and
allocation mechanisms that have been standardized
in EPONs, as well as with other popular well-
IPACT variations (Limited, Gated).
ACKNOWLEDGEMENTS
This work has been funded by the NSRF (2007-
2013) Synergasia-II/EPAN-II Program "Asymmetric
Passive Optical Network for xDSL and FTTH
Access," General Secretariat for Research and
Technology, Ministry of Education, Religious
Affairs, Culture and Sports (contract no. 09SYN-71-
839).
REFERENCES
B. Mukherjee, 2006. “Optical WDM Networks”, Springer,
University of California, Davis.
Jue, Jason P., Vokkarane, Vinod M., 2005. “Optical Burst
Switched Networks”, Optical Networks Series,
Springer.
G. Kramer, B. Mukherjee, G. Pesavento, 2002. "IPACT: a
dynamic protocol for Ethernet PON (EPON)", IEEE
Communications Magazine, vol. 40, no. 2, pp. 74–80.
Y. Luo, N. Ansari, 2005. "Limited sharing with traffic
prediction for dynamic bandwidth allocation and QoS
provisioning over EPONs," OSA J. Optical
Networking, vol. 4, no. 9, pp. 561–572.
G. Kramer, B. Mukherjee, and G. Pesavento, 2001.
“Ethernet PON (ePON): Design and Analysis of an
Optical Access Network,” Photonic Network
Communications, vol. 3, no. 3, pp. 307-319.
G. Kramer, B. Mukherjee, and A. Maislos, 2003.
“Ethernet Passive Optical Networks”, In S. Dixit
(Ed.), “Multiprotocol over DWDM: Building the Next
Generation Optical Internet”, John Wiley & Sons, pp.
229-260.
M. Mcgarry, M. Reisslein, M. Maier, 2008. "Ethernet
passive optical network architectures and dynamic
allocation algorithms," IEEE Communications
Surveys, vol. 3, no. 10, pp. 46-60.
N. Sadek, A. Khotanzad, 2004. “A dynamic bandwidth
allocation using a two-stage fuzzy neural network
based traffic predictor”, Proccedings of IEEE
International Conference on Neural Networks,
Hungary, pp. 2407-2412.
I.-S. Hwang, Z.-D. Shyu, L.-Y. Ke, C.-C. Chang, 2008.
“A novel early DBA mechanism with prediction-based
fair excessive bandwidth allocation scheme in EPON”,
Elsevier Computer Communications, vol. 31, pp.
1814–1823.
I.-S. Hwang, J-Y Lee, A. Liem, 2012. “QoS-based Genetic
Expression Programming Predicition Scheme in the
EPON’s”, Progress In Electromegnetics Research
Symposium Proceedings, 1589.
D. Swades, S. Vaibhav, G. M. Hari, S. Navrati, R.
Abhishek, 2010. “A new predictive dynamic priority
scheduling in Ethernet passive optical networks
(EPONs)”, Optical Switching and Networking, vol. 7,
pp. 215-223.
D. Morato, J. Acacil, L.A Diez, M. Izal, E. Magana, 2001.
“On linear prediction of Internet traffic for packet and
burst switching networks”, IEEE ICCN.
C.A. Chan, M. Attygalle, A. Nirmalathas, 2009. “Local
traffic prediction-based bandwidth allocation scheme
in EPON with active forwarding remote repeater node
” 14th OptoElectronics and Communications
Conference.
“OMNeT++ simulator”, http://www.omnetpp.org/
Z. Zhu, 2012. “Design of Energy-Saving Algorithms for
Hybrid Fiber Coaxial Networks Based on the DOCSIS
3.0 Standard”, IEEE/OSA Journal of Optical
Communications and Networking, vol. 4, pp. 449-
456.
OnthePredictionofEPONTrafficUsingPolynomialFittinginOpticalNetworkUnits
21
